Turn aimless AI evolution into goal-directed growth — guided by you, visualized as an RPG.
将 AI 的无方向进化变成有目标的成长 — 由你引导，以 RPG 形式呈现。
AI agents can run autonomously — but they can't decide what to get better at. AutoGPT, CrewAI, Devin — they execute tasks, but no one tells them "focus on Rust, not Python" or "stop wasting time on DevOps." The agent evolves, but without direction.
AI Agent 可以自主运行 — 但它们无法决定该往哪个方向变强。它们执行任务，但没人告诉它们"专攻 Rust，别学 Python"或"别在 DevOps 上浪费时间了"。Agent 在进化，但没有方向。
Hermes Quest turns you into the agent's guide. Define learning domains, give thumbs up/down on outcomes, and the agent adjusts its training direction — automatically, every cycle. It's prompt-level RLHF: your feedback doesn't retrain the model, it reshapes the agent's decision context. Same effect, zero cost.
Hermes Quest 让你成为 Agent 的引路人。定义学习领域、对结果点赞/差评，Agent 自动调整训练方向 — 每个周期都会。这是提示词级别的 RLHF：你的反馈不改模型权重，而是重塑 Agent 的决策上下文。相同效果，零成本。
The RPG layer isn't decoration — it lowers the cognitive cost of steering an AI. Instead of editing prompts and reading logs, you click a thumbs-down on a quest result, and the agent pivots. No technical background required.
RPG 层不是装饰 — 它降低了指挥 AI 的认知门槛。你不需要改 prompt、看日志，只需要对任务结果点个 👎，Agent 就会自动转向。
| Traditional Agent | Hermes Quest | |
|---|---|---|
| Evolution direction | Agent decides (or random) | User defines domains + feedback guides |
| Correction method | Rewrite prompts, restart | Click 👎, auto-adjusts next cycle |
| Observability | Read logs | RPG dashboard, 4-phase live progress |
| Cost to redirect | High (rewrite instructions) | Near-zero (one click) |
| User mental model | "I operate the AI" | "I guide the adventurer" |
Map / 地图
|
Guild / 公会
|
Shop / 商店
|
Tavern / 酒馆
|
Star Graph / 星图
|
Chronicle / 编年史
|
REFLECT -----> PLAN -----> EXECUTE -----> REPORT
(read feedback) (pick target) (train skill) (summarize)
^ |
+--- user 👍/👎 --- feedback-digest.json --+
- You define the world — name continents ("Machine Learning", "Music", "Rust") on the fog-of-war map
- Agent trains autonomously — each evolution cycle, it reflects, plans, executes, and reports (all 4 phases visible live)
- You steer with feedback — 👍 on good outcomes, 👎 on bad ones → aggregated into
feedback-digest.json - Agent adjusts next cycle — reads the digest at skill-level precision. Avoids specific skills you disliked, deepens skills you liked — without abandoning the entire domain
- Repeat — the agent gets better at what you care about, not what it randomly picks
We ran three real cycles to prove the loop closes:
| Cycle | Feedback State | Agent Decision | Result |
|---|---|---|---|
| #1 (baseline) | No feedback | Agent chose DevOps training | Autonomous default |
| #2 | DevOps 👎×3, Creative Arts 👍×4 | Pivoted to Creative Arts | "Pivoting away from DevOps due to 100% negative feedback" |
| #3 | DevOps/docker 👍×4, DevOps/ci-cd 👎×3, Creative 👍×4 | Stayed in DevOps but switched to Docker | "Positive feedback on docker-compose suggests deep-diving into container orchestration, despite negative signal on testing baseline" |
Cycle #3 is the key result: the agent didn't abandon DevOps entirely — it read the skill-level sentiment and pivoted within the domain. This is the difference between a blunt "avoid this area" and precise "avoid this approach, try another."
Click the fog → name a new continent → the agent starts learning that domain. Music? Cooking? Quantum Physics? You decide what your AI studies. Skills are automatically classified by LLM into the domains you define. Delete a domain, and skills redistribute.
点击迷雾 → 命名新大陆 → Agent 开始自学该领域。音乐？烹饪？量子物理？ 你来决定。
Not just a vanity metric. Your 👍/👎 flows into a structured feedback-digest.json — tracking sentiment per skill, per domain, with auto-generated corrections. The agent reads this at the start of every cycle and explicitly states how it's responding to your feedback.
不只是数字变化。你的 👍/👎 会流入结构化的反馈摘要 — 按技能、按领域追踪情感倾向。Agent 每个周期开始时读取，并明确说明如何回应你的反馈。
Every evolution cycle runs 4 observable phases: REFLECT → PLAN → EXECUTE → REPORT. The dashboard shows a live progress indicator — you see the agent analyzing its weaknesses, picking a target, training, and summarizing results. No more black-box "cycle complete" messages.
每个进化周期有 4 个可观测阶段。仪表盘实时显示进度 — 你能看到 Agent 在分析弱点、选择目标、训练、总结。不再是黑箱。
5 NPCs powered by real LLM — not scripted dialogue trees. Gus the bartender searches X/Twitter and retells it as tavern gossip. Orin the sage analyzes your actual game stats. The guild master recommends quests based on your weakest domains. They have group conversations too.
5 个 NPC 由真实 LLM 驱动 — 酒保搜 X 新闻，贤者分析真实数据，公会长推荐任务，还有群聊模式。
Everything runs through Hermes Agent's native extension points: Skills, Cron, Memory. No source code modifications. The dashboard is a pure observation + steering layer.
全部通过 Hermes Agent 原生扩展点运行。零源码修改。Dashboard 是纯粹的观测 + 引导层。
| Feature | Description |
|---|---|
| 🧙 Character Panel | HP (stability), MP (morale), XP, Gold, Level, Class, Title — all real-time |
| 🗺️ Custom World Map | 6 user-definable sites, fog-of-war, skill constellation star graphs |
| ⚔️ Guild Quest Board | AI-recommended quests targeting weak domains, custom creation, full lifecycle |
| 🔁 Feedback Loop | 👍/👎 → feedback-digest.json → agent reads next cycle → adjusts direction |
| 📊 Cycle Observability | REFLECT → PLAN → EXECUTE → REPORT — watch the agent think in real-time |
| 🍺 NPC Tavern | 5 LLM-powered NPCs: Lyra, Aldric, Kael, Gus, Orin — each with unique role |
| 💬 Group Chat | Multi-agent tavern chatter — NPCs discuss your adventures autonomously |
| 🏪 Skill Shop | Browse 80+ community skills from Hermes Hub, one-click install (300G) |
| 📰 Rumors Board | Real-time X/Twitter feed via Gus the bartender |
| 📜 Adventure Chronicle | Event timeline with feedback buttons — can mark quests as failed |
| 🧠 Skill Classification | LLM-powered: rename a domain → all skills auto-reclassified |
| 🎒 Inventory | Bag items with file viewer — research notes, training reports, code snippets |
| 🧪 Potions | HP Potion (200G) and MP Potion (150G) |
+-------------------+---------------------+-------------------+
| Quest Skill | FastAPI Backend | React Dashboard |
| (SKILL.md) | (Port 8420) | (Pixel RPG UI) |
| | | |
| - 4-phase cycle | - 40+ API endpoints | - Zustand store |
| - Reads digest | - feedback-digest | - WebSocket sync |
| - RPG rules | - /api/npc/chat | - Animated BGs |
| - XP/HP formulas | - /api/cycle/start | - 18+ panels |
| | - skill_classify | - cycle_progress |
+-------------------+---------------------+-------------------+
| Hermes Agent Runtime |
| Skills - Cron - Memory - Telegram - Hub |
+-------------------------------------------------------------+
| File | Location | Role |
|---|---|---|
state.json |
~/.hermes/quest/ |
Character stats — agent reads/writes, dashboard observes |
events.jsonl |
~/.hermes/quest/ |
Event stream — agent appends, watcher broadcasts |
feedback-digest.json |
~/.hermes/quest/ |
Aggregated user feedback — dashboard writes, agent reads |
knowledge-map.json |
~/.hermes/quest/ |
Skill domains, mastery levels, connections |
SKILL.md |
~/.hermes/skills/quest/ |
Agent behavior template (sync via API) |
- Hermes Agent installed
- Node.js 18+ and Python 3.11+
# Clone
git clone https://github.com/nemoaigc/hermes-quest.git
cd hermes-quest/hermes-quest-dashboard
# Configure environment
cp .env.example .env
# Edit .env with your API keys
# Frontend
npm install
npm run dev
# Backend (another terminal)
cd server
pip install -r requirements.txt
python -m uvicorn main:app --host 0.0.0.0 --port 8420# Sync SKILL.md template to Hermes (via API)
curl -X POST http://localhost:8420/api/skill/quest/sync
# Or manually
cp templates/quest-skill.md ~/.hermes/skills/quest/SKILL.mdnpm run build
# FastAPI serves dist/ automatically at port 8420| Stat | Meaning | Formula |
|---|---|---|
| HP | Stability (agent reliability) | 50 + level * 10, fail penalty: -15 |
| MP | Morale (user confidence) | 100 max, feedback: ±15, decay: -2/day |
| XP | Experience | level * 100 to next level |
| Gold | Currency | Quest rewards: 100-230G scaled by level |
Classes emerge from skill distribution: Warrior (coding), Mage (research), Ranger (automation), Paladin (balanced), Necromancer (delegation).
Economy: Quest create FREE, retry 50G, HP potion 200G, MP potion 150G, skill install 300G, board refresh 50G.
hermes-quest-dashboard/
+-- src/
| +-- panels/ # 18 UI panels (Map, Guild, Shop, Tavern...)
| +-- components/ # Shared (AnimatedBg, RpgButton, ErrorBoundary...)
| +-- constants/ # Theme, API config, NPC definitions
| +-- store.ts # Zustand state management
| +-- websocket.ts # WebSocket client + cycle_progress handler
| +-- api.ts # 20+ API client functions
| +-- types.ts # TypeScript interfaces
+-- server/
| +-- main.py # FastAPI app (40+ endpoints)
| +-- watcher.py # File watcher (2s poll) + event broadcasting
| +-- npc_chat.py # LLM-powered NPC dialogue
| +-- skill_classify.py # LLM skill-to-domain classification
| +-- ws_manager.py # WebSocket broadcast manager
+-- templates/
| +-- quest-skill.md # SKILL.md template (feedback rules + 4-phase cycle)
+-- public/ # Pixel art assets, sprites, fonts
- Feedback digest pipeline (👍/👎 →
feedback-digest.json→ agent behavior) - Cycle observability (4-phase progress: REFLECT → PLAN → EXECUTE → REPORT)
- End-to-end validation — 3 real cycles proving skill-level steering works
- SKILL.md v3.0 in NML with skill-level feedback precision
- Persistent feedback dedup via SQLite (
has_feedback_for_event) - Cycle progress recovery on reconnect (
_read_latest_cycle_progress) - Auto-sync SKILL.md before each cycle (
_sync_quest_skill_template) - Cycle lock auto-cleanup on report phase
- Subprocess log capture — replace DEVNULL with rotating log file for cycle debugging
- LLM compliance monitoring — post-cycle audit verifying agent followed SKILL.md rules
- MP morale awareness in SKILL.md prompt (low MP → safer quest choices)
- Workflow resolution improved —
_merge_event_contextmerges frontend data + event log lookup - Feedback-driven quest recommendations — filter by skill_sentiment, boost positives
- Feedback visibility — REFLECT reasoning shown in Chronicle, not just 40-char truncation
- Exploration mode — epsilon-greedy skill sampling to prevent filter bubbles
- Multi-dimensional feedback — 👎 followed by "wrong direction / poor quality / not relevant"
- Level-up celebration animation
- NPC awareness of recent feedback events
- Feedback time decay — old feedback loses weight via exponential decay
- Agent-to-human clarification channel — agent asks one question per cycle when uncertain
- Cycle lock staleness recovery — auto-clear if no phase events for 5+ minutes
- Achievement system — milestones, collections, hidden achievements with rewards
- MP-at-zero "burnout" event — triggers recovery quest
-
pip install hermes-quest-dashboard— ship as Python package with bundled frontend - New player tutorial — first-run guided setup
- Skill search/filter (by category, rarity, source)
- Agent hypothesis generation (analyze why failures happen)
- Forgetting curves + confidence intervals for mastery
- YAML/CLI config export — lightweight alternative for power users (Butterfield)
- Tauri desktop app (DMG) with PyInstaller sidecar
MIT — Use it, fork it, evolve it.
Built by Nemoverse | Powered by Hermes Agent
You define the world. The agent conquers it.